DeepSeek V4-Pro goes GA at 1.65T, MIT — and quietly raises its prices on Saturday

DeepSeek published DeepSeek-V4-Pro-0813 today, superseding the April preview: 1.65 trillion parameters, MIT licence, 1M context. Coming a week after Qwen opened a 2.4T model under a custom licence and a fortnight after Kimi K3 opened 2.78T under a custom licence, this is now the largest open-weight model available under a plain OSI licence. It also comes with a pricing change most coverage will miss.

Verdict: The largest open-weight model under a classic OSI licence — and a rate card that changes in three days

The take

The release, verified against the Hugging Face API and DeepSeek's own rate card: `deepseek-ai/DeepSeek-V4-Pro-0813`, created August 13 2026, MIT, safetensors totalling 1,650,497,936,906 parameters across 92 files. Architecturally it is the preview's structure — 43 layers, 256 routed experts plus one shared, 6 active per token, FP4 expert weights — with a DSpark speculative-decoding module attached at layers 40–42. DeepSeek reports large agentic gains over the preview: Terminal Bench 2.1 72.1 → 87.9, DeepSWE 12.8 → 62.7, NL2Repo 38.5 → 61.5, Cybergym 52.7 → 83.3. On those rows it sits level with Kimi K3 and ahead of Opus 4.8. All of it is vendor-run on DeepSeek's own harness at max reasoning effort, and two of the ten benchmarks are internal test sets they alone can run — so treat the table as a claim, not a finding, until someone independent reproduces it.

**The pricing change is the part worth acting on.** DeepSeek now publishes a rate card for both V4 models, which earlier sweeps of ours could not find at all: `deepseek-v4-pro` at $0.435 per 1M input on a cache miss and $0.87 output; `deepseek-v4-flash` at $0.14 / $0.28. But the same page states that at **16:00 UTC on August 16, 2026** — three days from publication — billing moves to peak/off-peak. V4-Pro goes to $0.66 / $1.98 off-peak and $1.32 / $3.96 at peak, where peak is 01:00–04:00 and 06:00–10:00 UTC. Read those numbers carefully: the new **off-peak** rate is higher than today's flat rate. This is a price increase wearing the clothing of a discount structure, and V4-Pro roughly doubles at best and quadruples at worst.

Our call: model entry updated at `/models/deepseek-v4-pro/`, no planner pick, and deliberately **no cost-calculator row** — for either V4 model. We could publish today's flat rate honestly, and it would be wrong by Saturday. The calculator exists so people can reason about steady-state monthly spend, and a two-tier time-of-day rate cannot be reduced to one blended number without inventing a usage distribution we do not have. Better to say so than to ship a figure with a three-day shelf life. We will add both rows after the new structure has settled and we can see what people actually pay.

The licence context is what makes this release matter beyond DeepSeek. Three of the largest open-weight models in history landed inside four weeks — Kimi K3 (2.78T, custom `kimi-k3` licence), Qwen3.8-Max (2.4T, custom `qwen3.8-max` licence with revenue clauses), and this one. Only DeepSeek's is MIT. If you are choosing an open frontier model on licence clarity rather than on benchmark position, the field is smaller than the headlines suggest. None of the three is remotely local: at 1.65T this needs a datacenter, and the honest local path from the V4 line is still V4-Flash at ~158 GB Q4, which means an M3 Ultra or dual 80 GB server cards.

Where this fits

Models: DeepSeek V4-Pro · DeepSeek V4-Flash · Kimi K3 (2.8T-A50B) · Qwen3.8-Max (2.4T-A95B)

Hardware: NVIDIA DGX Spark · Mac Studio M3 Ultra 96 GB · Dual RTX 5090

Sources

Next step

Try this in the planner→